Flemish Dutch Pre-training Volume for Few-Shot Noisy Speech Adaptation in Self-Supervised Models

Recent research in speech processing exhibits a growing interest in unsupervised and self-supervised representation learning from unlabelled data to alleviate the need for large amounts of annotated data. We investigate several popular pre-training methods and apply them to Flemish Dutch. We compare off-the-shelf English pre-trained models to models trained on an increasing amount of Flemish data. We find that the most important factors for positive transfer to downstream speech recognition tasks include a substantial amount of data and a matching pre-training domain. Ideally, we also finetune

Research goal: Does increasing the volume of Flemish Dutch pre-training data improve the few-shot adaptation accuracy of self-supervised models on noisy speech subsets of CommonVoice compared to English-only pre-trained baselines?

Autonomous synthesis report generated by Assignee Research. Tribunal consensus score: 7.8/10.